Meta scrambles to find out why DeepSeek’s AI is so good

Mark Zuckerberg’s Meta has urgently assembled as many as four “war rooms” to determine how DeepSeek, a modest Chinese AI startup, managed to release an AI assistant, already called a game-changer.

DeepSeek released a version of its AI chatbot, R1, that it says can perform just as well as rival models, such as ChatGPT, at a fraction of the cost. Besides building a model that allegedly required much less computing power – it’s also open-source.

“Given their open-source availability and influence, it's reasonable to assume that Meta's Llama models provided a blueprint or inspiration for aspects of DeepSeek's development,” technologist Jason Snyder wrote in Forbes this week.

However, Meta undoubtedly didn’t expect DeepSeek's cheap model to be so good or even better than the American tech giant’s new systems.

According to The Information, Meta AI infrastructure director Mathew Oldham has told colleagues that DeepSeek’s newest model could outperform even the next version of Meta’s Llama AI.

Zuckerberg previously said the latter could be released in early 2025 – which must be very soon indeed. That’s why Meta is now scrambling to try and keep up with DeepSeek.

Of the four war rooms Meta has reportedly created to respond to DeepSeek’s rise, two teams will try to determine how High-Flyer Capital Management, a hedge fund backing DeepSeek, lowered the cost of training and running the new AI assistant.

The goal is, of course, to adapt those tactics for Llama, an anonymous Meta employee told The Information.

Another team will try to find out what data DeepSeek used to train its model, and the fourth one will consider how Llama can restructure its models based on attributes of the DeepSeek models.

Meta CEO hasn’t yet directly responded to DeepSeek’s rise, but he said on Facebook on Friday that a new Llama model version would become “the leading state-of-the-art model” upon release. He also announced that the company would spend as much as $65 billion on AI projects in 2025.

The company’s chief AI scientist, Yann LeCun, said on LinkedIn that he’s calm. In a post, he said: “To people who see the performance of DeepSeek and think: ‘China is surpassing the US in AI.’ You are reading this wrong. The correct reading is: ‘Open source models are surpassing proprietary ones.’”

“They [DeepSeek] came up with new ideas and built them on top of other people’s work. Because their work is published and open source, everyone can profit from it. That is the power of open research and open source,” said LeCun.